S3: Size-aware Sequential Scheduling to Meet Deadlines in Data Center Networks

S3: Size-aware Sequential Scheduling to Meet Deadlines in Data Center Networks

In modern Data Center Network (DCN), there exist various on-line interactive application services, such as web search, social networking and online transaction processing. An urgent demand for these on-line interactive applications is low deadline miss ratio. Though the flows with small or tiny size are overwhelming for most data center applications, previous deadline-aware transport protocols have no consideration for protection of these flows, with the inevitable result that few large flows occupy too much network service time and many small flows could not meet their deadlines. Therefore, the deadline miss ratio may be improved further if the flow sizes are considered by transport protocols. In this paper, we first analyze the distribution of the flow size according to the real DCN traffic trace. Then, based on the analysis, we propose a preemptive distribution flow scheduling protocol, Size-aware Sequential Scheduling (S3 ), which schedules flows such that they can send data at their maximal sending rates and finish as quickly as possible. More importantly, different from the previous transport protocol studies, S3 sequentially adjusts the flow scheduling order with the consideration of both the flow size and the flow deadline, which can help more small flows meet their deadlines. The at-scale simulations and a real flow distribution based test results show that S3 significantly reduces the deadline miss ratio and obtains lower flow completion time compared to the recent protocols PDQ and D3 . In addition, S3 is resilient to packet loss, and schedules flows with only the same time complexity as PDQ.